File size: 8,176 Bytes
0441b41 b2b5d5f 0441b41 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 |
from concurrent.futures import ThreadPoolExecutor, as_completed
import json
import os
import time
import numpy as np
import requests
import torch
from clip_app_client import ClipAppClient
from clip_retrieval.clip_client import ClipClient, Modality
clip_retrieval_service_url = "https://knn.laion.ai/knn-service"
map_clip_to_clip_retreval = {
"ViT-L/14": "laion5B-L-14",
"open_clip:ViT-H-14": "laion5B-H-14",
"open_clip:ViT-L-14": "laion5B-L-14",
}
def safe_url(url):
import urllib.parse
url = urllib.parse.quote(url, safe=':/')
# if url has two .jpg filenames, take the first one
if url.count('.jpg') > 0:
url = url.split('.jpg')[0] + '.jpg'
return url
# test_image_path = os.path.join(os.getcwd(), "images", "plant-001.png")
test_image_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "images", "plant-001.jpeg")
# test_image_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "images", "plant-002.jpeg")
# test_image_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "images", "plant-002.jpeg")
# test_image_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "images", "car-002.jpeg")
app_client = ClipAppClient()
clip_retrieval_client = ClipClient(
url=clip_retrieval_service_url,
indice_name=map_clip_to_clip_retreval[app_client.clip_model],
# use_safety_model = False,
# use_violence_detector = False,
# use_mclip = False,
num_images = 300,
# modality = Modality.TEXT,
# modality = Modality.TEXT,
)
preprocessed_image = app_client.preprocess_image(test_image_path)
preprocessed_image_embeddings = app_client.preprocessed_image_to_embedding(preprocessed_image)
print (f"embeddings: {preprocessed_image_embeddings.shape}")
embedding_as_list = preprocessed_image_embeddings[0].tolist()
results = clip_retrieval_client.query(embedding_input=embedding_as_list)
# hints = ""
# for result in results:
# url = safe_url(result["url"])
# similarty = float("{:.4f}".format(result["similarity"]))
# title = result["caption"]
# print (f"{similarty} \"{title}\" {url}")
# if len(hints) > 0:
# hints += f", \"{title}\""
# else:
# hints += f"\"{title}\""
# print("---")
# print(hints)
image_labels = [r['caption'] for r in results]
image_label_vectors = [app_client.text_to_embedding(label) for label in image_labels]
image_label_vectors = torch.cat(image_label_vectors, dim=0)
dot_product = torch.mm(image_label_vectors, preprocessed_image_embeddings.T)
similarity_image_label = [(float("{:.4f}".format(dot_product[i][0])), image_labels[i]) for i in range(len(image_labels))]
similarity_image_label.sort(reverse=True)
for similarity, image_label in similarity_image_label:
print (f"{similarity} {image_label}")
print (f"----\n")
# now do the same for images
def _safe_image_url_to_embedding(url, safe_return):
try:
return app_client.image_url_to_embedding(url)
except:
return safe_return
image_urls = [safe_url(r['url']) for r in results]
image_vectors = [_safe_image_url_to_embedding(url, preprocessed_image_embeddings * 0) for url in image_urls]
image_vectors = torch.cat(image_vectors, dim=0)
dot_product = torch.mm(image_vectors, preprocessed_image_embeddings.T)
similarity_image = [(float("{:.4f}".format(dot_product[i][0])), image_labels[i]) for i in range(len(image_labels))]
similarity_image.sort(reverse=True)
for similarity, image_label in similarity_image:
print (f"{similarity} {image_label}")
def mean_template(embeddings):
template = torch.mean(embeddings, dim=0, keepdim=True)
return template
def principal_component_analysis_template(embeddings):
mean = torch.mean(embeddings, dim=0)
embeddings_centered = embeddings - mean # Subtract the mean
u, s, v = torch.svd(embeddings_centered) # Perform SVD
template = u[:, 0] # The first column of u gives the first principal component
return template
def clustering_templates(embeddings, n_clusters=5):
from sklearn.cluster import KMeans
import numpy as np
kmeans = KMeans(n_clusters=n_clusters)
embeddings_np = embeddings.numpy() # Convert to numpy
clusters = kmeans.fit_predict(embeddings_np)
templates = []
for cluster in np.unique(clusters):
cluster_mean = np.mean(embeddings_np[clusters == cluster], axis=0)
templates.append(torch.from_numpy(cluster_mean)) # Convert back to tensor
return templates
# create a templates using clustering
print(f"create a templates using clustering")
merged_embeddings = torch.cat([image_label_vectors, image_vectors], dim=0)
clusters = clustering_templates(merged_embeddings, n_clusters=5)
# convert from list to 2d matrix
clusters = torch.stack(clusters, dim=0)
dot_product = torch.mm(clusters, preprocessed_image_embeddings.T)
cluster_similarity = [(float("{:.4f}".format(dot_product[i][0])), i) for i in range(len(clusters))]
cluster_similarity.sort(reverse=True)
for similarity, idx in cluster_similarity:
print (f"{similarity} {idx}")
# template = highest scoring cluster
# template = clusters[cluster_similarity[0][1]]
template = preprocessed_image_embeddings * (len(clusters)-1)
for i in range(1, len(clusters)):
template -= clusters[cluster_similarity[i][1]]
print("---")
print(f"seaching based on template")
results = clip_retrieval_client.query(embedding_input=template[0].tolist())
hints = ""
for result in results:
url = safe_url(result["url"])
similarty = float("{:.4f}".format(result["similarity"]))
title = result["caption"]
print (f"{similarty} \"{title}\" {url}")
if len(hints) > 0:
hints += f", \"{title}\""
else:
hints += f"\"{title}\""
print(hints)
# cluster_num = 1
# for template in clusters:
# print("---")
# print(f"cluster {cluster_num} of {len(clusters)}")
# results = clip_retrieval_client.query(embedding_input=template.tolist())
# hints = ""
# for result in results:
# url = safe_url(result["url"])
# similarty = float("{:.4f}".format(result["similarity"]))
# title = result["caption"]
# print (f"{similarty} \"{title}\" {url}")
# if len(hints) > 0:
# hints += f", \"{title}\""
# else:
# hints += f"\"{title}\""
# print(hints)
# cluster_num += 1
# create a template
# mean
# image_label_template = mean_template(image_label_vectors)
# image_template = mean_template(image_vectors)
# pca
# image_label_template = principal_component_analysis_template(image_label_vectors)
# image_template = principal_component_analysis_template(image_vectors)
# clustering
# image_label_template = clustering_template(image_label_vectors)
# image_template = clustering_template(image_vectors)
# take the embedding and subtract the template
# image_label_template = preprocessed_image_embeddings - image_label_template
# image_template = preprocessed_image_embeddings - image_template
# image_label_template = image_label_template - preprocessed_image_embeddings
# image_template = image_template - preprocessed_image_embeddings
# normalize
# image_label_template = image_label_template / image_label_template.norm()
# image_template = image_template / image_template.norm()
# results = clip_retrieval_client.query(embedding_input=image_label_template[0].tolist())
# hints = ""
# print("---")
# print("average of image labels")
# for result in results:
# url = safe_url(result["url"])
# similarty = float("{:.4f}".format(result["similarity"]))
# title = result["caption"]
# print (f"{similarty} \"{title}\" {url}")
# if len(hints) > 0:
# hints += f", \"{title}\""
# else:
# hints += f"\"{title}\""
# print(hints)
# print("---")
# print("average of images")
# results = clip_retrieval_client.query(embedding_input=image_template[0].tolist())
# hints = ""
# for result in results:
# url = safe_url(result["url"])
# similarty = float("{:.4f}".format(result["similarity"]))
# title = result["caption"]
# print (f"{similarty} \"{title}\" {url}")
# if len(hints) > 0:
# hints += f", \"{title}\""
# else:
# hints += f"\"{title}\""
# print(hints) |